Reinforcement Learning Approaches to the Dynamic Flexible Job Shop Scheduling Problem: A Systematic Review

This systematic review analyses how Reinforcement Learning (RL) applied to the Dynamic Flexible Job Shop Scheduling Problem (DFJSSP) contributes to industrial production systems, considering problem formulations and extensions, algorithm families and scheduling strategies, and scalability. ScienceDirect, IEEE Xplore, Taylor & Francis and Web of Science databases were searched between November 2025 and January 2026, using eligibility criteria requiring RL as the primary method, explicit modelling of dynamic events, experimental validation, and quantitative instance data. A total of 231 records were identified, from which 43 studies were included. Results show that 21 studies incorporate additional resources, such as transport and energy and human factors. The review shows that PPO and DQN approaches account for 37% and 35% of studies respectively, with single-agent architectures prevailing in 65% of cases. It was identified that most studies focus on small-scale to medium-scale simulated instances, and heterogeneous parameterisation of dynamic events limits cross-study comparability. In addition, this review characterises the sustainability dimensions (economic, environmental and social) addressed by RL approaches to the DFJSSP, based on the optimisation objectives adopted in each study. Future research should jointly address economic, environmental and social objectives with explicit indicators and validate RL on real industrial data to establish it as an enabler of sustainable manufacturing in DFJSSP.